Embed Size (px)
Citation preview
San Jose State UniversitySJSU ScholarWorks
Master's Projects Master's Theses and Graduate Research
Spring 5-25-2016
Malicious JavaScript Detection using StatisticalLanguage ModelAnumeha ShahSan Jose State University
Follow this and additional works at: https://scholarworks.sjsu.edu/etd_projects
Part of the Information Security Commons
This Master's Project is brought to you for free and open access by the Master's Theses and Graduate Research at SJSU ScholarWorks. It has beenaccepted for inclusion in Master's Projects by an authorized administrator of SJSU ScholarWorks. For more information, please [email protected].
Recommended CitationShah, Anumeha, "Malicious JavaScript Detection using Statistical Language Model" (2016). Master's Projects. 476.DOI: https://doi.org/10.31979/etd.nujz-hf4ahttps://scholarworks.sjsu.edu/etd_projects/476
Malicious JavaScript Detection using Statistical Language Model
A Project
Presented to
The Faculty of the Department of Computer Science
San Jose State University
In Partial Fulfillment
of the Requirements for the Degree
Master of Science
by
Anumeha Shah
May 2016
c○ 2016
Anumeha Shah
ALL RIGHTS RESERVED
The Designated Project Committee Approves the Project Titled
Malicious JavaScript Detection using Statistical Language Model
by
Anumeha Shah
APPROVED FOR THE DEPARTMENTS OF COMPUTER SCIENCE
SAN JOSE STATE UNIVERSITY
May 2016
Dr. Thomas Austin Department of Computer Science
Dr. Chris Pollett Department of Computer Science
Dr. Jon Pearce Department of Computer Science
ABSTRACT
Malicious JavaScript Detection using Statistical Language Model
by Anumeha Shah
The Internet has an immense importance in our day to day life, but at the
same time, it has become the medium of infecting computers, attacking users, and
distributing malicious code. As JavaScript is the principal language of client side pro-
gramming, it is frequently used in conducting such attacks. Various approaches have
been made to overcome the JavaScript security issues. Some advanced approaches
utilize machine learning technology in combination with de-obfuscation and emula-
tion. Many methods of analysis incorporate static analysis and dynamic analysis.
Our solution is entirely based on static analysis, which avoids unnecessary runtime
overhead.
The central objective of this project is to integrate the work done by Eunjin (EJ)
Jung et al. on Towards A Robust Detection of Malicious JavaScript (TARDIS) into
the web browser via a Firefox add-on and to demonstrate the usability of our add-
on in defending against such attacks. TARDIS uses statistical language modeling
for an automatic feature extraction and combines it with structural features from
an abstract syntax tree [1]. We have developed a Firefox add-on that is capable
of extracting JavaScript code from the page visited and classifying the JavaScript
code as either malicious or benign. We leverage the benefit of using a pre-compiled
training model in JavaScript Object Notation (JSON). JSON is lightweight and does
not consume much memory on a user’s machine. Moreover, it stores the data as
key-value pairs and easily maps to the data structures used in modern programming
languages. The principle advantage of using a pre-compiled training model is better
performance. Our model can achieve 98% accuracy on our sample dataset.
ACKNOWLEDGMENTS
I would like to express my sincere gratitude to my thesis advisor, Dr. Thomas
Austin for his continuous support, valuable comments, and guidance throughout this
project. I would also like to thank my committee members, Dr. Chris Pollett and Dr.
Jon Pearce for their valuable time, and feedback. Very special thanks to Professor
Eunjin (EJ) Jung, for helping me in understanding the project, and providing me the
malicious dataset and the necessary platform to carry out the project execution.
v
Contents
Chapter
1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Our approach to the problem . . . . . . . . . . . . . . . . . . . . 2
1.2 Firefox add-on . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1 Cross site scripting (XSS) . . . . . . . . . . . . . . . . . . . . . . 5
2.1.1 Stored Cross Site Scripting . . . . . . . . . . . . . . . . . . 6
2.1.2 Reflected cross-site scripting . . . . . . . . . . . . . . . . . 7
2.1.3 DOM based XSS Attack . . . . . . . . . . . . . . . . . . . 8
2.2 Other variants of JavaScript Attack . . . . . . . . . . . . . . . . . 9
2.3 Security measures adopted to prevent malicious JavaScript Attack 11
2.4 Static Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.5 Dynamic Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.6 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.6.1 JStill (Mostly Static Approach) . . . . . . . . . . . . . . . 13
2.6.2 Zozzle: Fast and Precise In-Browser JavaScript MalwareDetection . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.6.3 Cujo: efficient detection and prevention of drive-by-download attacks . . . . . . . . . . . . . . . . . . . . . 14
2.6.4 EarlyBird: Early Detection of Malicious Behavior in JavaS-cript Code . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.6.5 Prophiler: A Fast Filter for the Large-Scale Detection ofMalicious Web Pages . . . . . . . . . . . . . . . . . . 15
vi
vii
2.6.6 Wepawet . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.6.7 PJScan: Static Detection of Malicious JavaScript-BearingPDF Documents [4] . . . . . . . . . . . . . . . . . . . . 16
2.6.8 IceShield: Detection and Mitigation of Malicious Websiteswith a Frozen DOM . . . . . . . . . . . . . . . . . . . 17
2.7 TARDIS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7.1 Abstract syntax tree . . . . . . . . . . . . . . . . . . . . . 22
2.7.2 Statistical Language Modeling (SLM) . . . . . . . . . . . 22
2.7.3 TARDIS SLM model . . . . . . . . . . . . . . . . . . . . . 23
2.7.4 N-grams SLM model . . . . . . . . . . . . . . . . . . . . . 23
2.7.5 Character level n-grams . . . . . . . . . . . . . . . . . . . 24
2.7.6 Keyword Transformation . . . . . . . . . . . . . . . . . . . 25
2.7.7 Composite word-type transformation . . . . . . . . . . . . 26
2.8 Malicious Probability Query Strategy . . . . . . . . . . . . . . . . 27
3 Firefox add-on Implementation . . . . . . . . . . . . . . . . . . . 28
3.1 Usability of our Firefox add-on . . . . . . . . . . . . . . . . . . . . 28
3.2 Developing a Firefox add-on . . . . . . . . . . . . . . . . . . . . . 28
3.3 WebExtensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.4 Add-on SDK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.5 Firefox Add-on SDK installation and structure . . . . . . . . . . . 29
3.6 index.js . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.7 Content scripts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.8 Data Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.8.1 SLM_Script.js . . . . . . . . . . . . . . . . . . . . . . . . . 34
viii
3.8.2 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.9 Pre-compiled training model . . . . . . . . . . . . . . . . . . . . . 35
3.9.1 Types of pre-compiled models . . . . . . . . . . . . . . . . 35
3.9.2 Character level n-gram model . . . . . . . . . . . . . . . . 36
3.9.3 Keyword transformation . . . . . . . . . . . . . . . . . . . 37
3.9.4 Composite word type transformation . . . . . . . . . . . . 38
3.9.5 Precompiled training models computation . . . . . . . . . 39
3.9.6 Problems faced during pre-compiled model generation andsolution implementation . . . . . . . . . . . . . . . . . 40
3.10 Firefox add-on implementation . . . . . . . . . . . . . . . . . . . . 43
3.11 Result Computation . . . . . . . . . . . . . . . . . . . . . . . . . 44
4 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1 Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.1 Malicious scripts . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.2 Benign Scripts . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.1.3 Problems with the scripts . . . . . . . . . . . . . . . . . . 46
4.2 Training models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.3 Evaluation of n-grams models . . . . . . . . . . . . . . . . . . . . 46
4.3.1 Evaluation of optimized and non-optimized models of char-acter level n-grams. . . . . . . . . . . . . . . . . . . . . 47
4.3.2 Keyword transformation . . . . . . . . . . . . . . . . . . . 48
4.3.3 Composite word type transformation . . . . . . . . . . . . 49
4.4 Model comparisons regarding accuracy and detection time onsample data set. . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
ix
List of Tables
1 Performance comparison of optimized and non-optimized characterlevel n-grams model. . . . . . . . . . . . . . . . . . . . . . . . . . 47
2 Accuracy and precision evaluation of optimized and non-optimizedcharacter level n-grams model. . . . . . . . . . . . . . . . . . . . . 47
3 Performance comparison of optimized and non-optimized Keywordtransformation n-grams model. . . . . . . . . . . . . . . . . . . . 48
4 Accuracy and precision evaluation of optimized and non-optimizedkeyword transformation level n-grams model. . . . . . . . . . . . 48
5 Evaluation of composite word type transformation n-grams model. 49
6 Accuracy and precision evaluation of composite word type trans-formation n-grams model. . . . . . . . . . . . . . . . . . . . . . . 49
x
List of Figures
1 Stored cross-site scripting attack . . . . . . . . . . . . . . . . . . 6
2 Reflected cross-site scripting attack . . . . . . . . . . . . . . . . . 7
3 DOM based cross-site scripting attack . . . . . . . . . . . . . . . 8
4 DOM based cross-site scripting attack example . . . . . . . . . . 8
5 DOM based cross-site scripting attack example . . . . . . . . . . 9
6 HTML page with embedded malicious JavaScript . . . . . . . . . 9
7 Cross-site request frogery attack . . . . . . . . . . . . . . . . . . . 10
8 Malicious request forged by the attacker . . . . . . . . . . . . . . 10
9 Benign script sample . . . . . . . . . . . . . . . . . . . . . . . . . 18
10 Malicious script sample . . . . . . . . . . . . . . . . . . . . . . . . 19
11 Obfuscated script sample . . . . . . . . . . . . . . . . . . . . . . . 20
12 Initial directory structure of the Firefox add-on . . . . . . . . . . 30
13 Index.js . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
14 function runScript.js . . . . . . . . . . . . . . . . . . . . . . . . . 31
15 Add-on directory structure . . . . . . . . . . . . . . . . . . . . . . 33
16 Example of port.emit . . . . . . . . . . . . . . . . . . . . . . . . . 34
17 A snapshot of a pre-compiled malicious character level n-grams model 36
18 a snapshot of a keyword transformation n-grams model . . . . . . 37
19 A snapshot of the malicious n-grams composite word type trans-formation model . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
20 Java code added for model computation . . . . . . . . . . . . . . 40
xi
xii
21 CPU utilization before multithreading implementation . . . . . . 41
22 CPU idle time before multithreading implementation . . . . . . . 41
23 CPU utilization after multithreading implementation . . . . . . . 42
24 CPU idle time after multithreading implementation . . . . . . . . 42
25 Firefox add-on in the browser . . . . . . . . . . . . . . . . . . . . 43
26 Firefox add-on detection result in the console . . . . . . . . . . . 43
27 Total number of words in benign keyword transform model . . . . 49
28 Total number of words in malicious keyword transform model . . 49
29 Accuracy comparison of all the three models . . . . . . . . . . . . 50
30 Performance comparison of all the three models . . . . . . . . . . 51
31 Performance comparisons of all the three models in real word scen-ario for the top websites . . . . . . . . . . . . . . . . . . . . . . . 52
CHAPTER 1
Introduction
JavaScript and its frameworks are popular choice among web developers for build-
ing web pages. JavaScript can be placed in the HTML of web pages and can inter-
face with the document object model of the page to provide extensive functionalities
such as form validation, animation, asynchronous behavior, user activity tracking,
interactivity, and more [9]. JavaScript is also used in server side code and in mo-
bile applications by using cross-platform development tools such as Titanium and
PhoneGap [10].
Since the release of JavaScript in 1995, many browsers and client-side security
issues which have gained widespread attention [9]. JavaScript’s capability to interact
with the page’s document object model makes it powerful, but at the same time,
it also opens doors for attackers who can run malicious scripts on client computers
by enabling a malicious agent to deliver the scripts over the internet. Malicious
JavaScript has been listed in the Open Web Application Security Project (OWASP)’s
2013 Top 10 List of security issues [2]. Cross-site scripting has been listed as the third
most widespread web application vulnerabilities on the Internet. Malicious JavaScript
payload can be embedded into a legitimate website or web application by an attacker
and can be executed on a client’s machine. Several security measures have been taken
to restrict the malicious code in order to access the client side sensitive information,
the malicious JavaScript has access to the same objects as web pages and includes
the user’s cookies, sessions, etc. The malicious code can also redirect a user to an
attacker’s website and execute some malicious code without the user’s permission,
further advancing the attack to more severe ones.
1
One approach to solving this problem is to identify the pages that contain ma-
licious scripts and either warn users before loading the page or block those scripts.
The problem arises is how to distinguish malicious scripts from the benign ones ac-
curately, as the dynamic nature of JavaScript makes it difficult to detect the exploit
code. Moreover, attackers often use sophisticated obfuscation techniques that hide
the malicious code and make detection complicated.
Recent work involves using machine learning techniques in combination with
de-obfuscation and emulation technology [1]. Machine learning is used for feature
extraction to identify the nature of the scripts. However, the malicious code keeps
evolving, taking benefits of the dynamic feature of JavaScript though, they still need
primitive JavaScript operations to be converted to clear text before execution [9]. A
machine learning combined with de-obfuscation/emulation has proved to be advant-
ageous, but they need a customized browser [1].
1.1 Our approach to the problem
Our approach is based on TARDIS [1]. TARDIS only requires the source code and
does not utilize any de-obfuscation techniques on the original source code. TARDIS
is simple yet achieves high accuracy compared to related research [1]. TARDIS uses
machine learning techniques and robust features. Robust features are the features
that can classify the malicious code with a high degree of accuracy. An attempt to
conceal these features in the malicious code will require modifications in the malicious
code generation algorithm, and to incorporate these modifications, an attacker will
require additional resources.
The intuition on which TARDIS is based is the difference in the utilization of
the JavaScript language for writing a benign program versus writing a malicious one.
2
An attacker writing malicious code attempts to conceal what the code is doing using
various automated or manual procedures and involves the use of regular expressions,
rules, or machine learning. A malicious program is likely to include more redundant
parts as compared to a benign program. A benign, but poorly written JavaScript pro-
gram may also include redundancy and inefficiency. However, an attacker’s intention
of bypassing the detection of the malicious code and the use of automation to generate
obfuscated script tend to include much more redundancy and inefficiency as compared
to a benign JavaScript program. TARDIS makes use of this difference. Furthermore,
the features have been extended with a Statistical Language Model (SLM). SLM is
termed as a probability distribution(s) of String S and estimates the frequency of a
String S in a sentence [11]. SLM uses the general patterns in the language used in
both malicious and benign JavaScript to classify benign and malicious JavaScript [1].
1.2 Firefox add-on
We have developed a Firefox add-on based on TARDIS. Once added to the
browser, this add-on is capable of capturing the inline JavaScript from the current
open tab. It then extracts the required features, performs analysis, and identifies
the existence of an exploit. On detection of malicious JavaScript, the Firefox add-on
alerts the user of the presence of an exploit in the current tab
Our Firefox add-on uses a precompiled training model in order to perform an
efficient prediction. The precompiled training model has been stored in JSON. JSON
is lightweight and allows a quick search. The training model has been computed over
15000 malicious and 30000 benign JavaScript files, and the model has been tested
using more than 1000 malicious and 1000 benign JavaScript files. A 10-fold cross-
validation has been performed in order to validate the model. The model tends to
3
reach 98% accuracy.
The remaining of the paper is organized as follows. In Chapter 2 we provide
background information on SLM, XSS, and discuss TARDIS and other related work.
Chapter 3 presents the Firefox add-on development and pre-compiled training model
and similar security research by top companies and universities. In Chapter 4 we
provide test results and accuracy of the training model, and Chapter 5 covers the
conclusion. tradeoffs, and future work.
4
CHAPTER 2
Background
JavaScript is one of the primary languages in programming web technologies.
It can interface with the document’s object model (DOM) and provides different
impressive functionality. Because of these features, JavaScript is extensively used on
nearly every website, and all of the browsers allow JavaScript, as it helps in making
the page dynamic and it keeps a user engaged.
JavaScript’s capability of interacting with the DOM also grants it with the po-
tential of injecting malicious code in the script dynamically. There has been various
flavors and types of malicious JavaScript, and one of the most wicked ones is cross-site
scripting (XSS).
2.1 Cross site scripting (XSS)
An XSS attack targets web applications that do not validate and sanitize user
input such as form data, comments, etc. in a proper way; that enables attackers to
inject malicious code into the web page. An attacker may insert a link to the third
party malicious website into the benign web page. If a user visits such an infected
page and clicks the link, the link will take the user to the malicious website and steal
the user’s cookies and other sensitive information stored in the browser. An attacker
can use this information to impersonate that user. Attackers can also employ various
kind of obfuscation technique to conceal the exploit in the link and makes it resemble
like a legitimate link. There are commonly three types of XSS attacks: stored XSS,
reflected XSS and DOM-based XSS [12].
5
2.1.1 Stored Cross Site Scripting
Stored cross-site scripting targets the websites that store the user input such as
comments, form, etc. first in databases or the file system and later requested by the
website users. If the input has not been sanitized or encoded and the data contains an
attack, The user will receive the malicious script. This type of attack affects multiple
users of the website [12].
Stored cross site scripting attack: attacker is storing malicious script to database
using a form. The data is stored in the database without proper input validation and
returned to the web user without output validation. A user clicks on the malicious
link and the attacker hijacks the information stored in user’s browser.
Figure 1: Stored cross-site scripting attack [14]
6
2.1.2 Reflected cross-site scripting
Reflected cross-site scripting targets the websites that reflect a user’s input im-
mediately to the web page. If not encoded, it may allow an attacker to introduce
malicious code into the dynamic webpage. However, an attacker can only change his
web page result, though the attacker can persuade a user to click on a link, which
can lead that user to a malicious website [13].
Reflected cross site scripting attack: an attacker identifies a vulnerable website
and inject malicious link. The attacker then convinces the user to click on the link
using social engineering. The user clicks on the link and becomes victim of reflected
XSS attack.
Figure 2: Reflected cross-site scripting attack [13]
7
2.1.3 DOM based XSS Attack
Every HTML page has an associated document object model (DOM) that con-
sists of the HTML page objects. These objects represent the document properties.
When a JavaScript within an HTML page executed, the browser provides the DOM
of the HTML page to the script. A JavaScript can interact with the DOM and may
perform an action based on the properties of the objects in DOM to make the page
more interactive and dynamic. A DOM XSS attack targets the improper treatment
of the data from its associated DOM in the HTML pages [20].
An example of DOM based XSS attack.
Figure 3: In this html page, JavaScript variable pos is set to the value of context fieldform the URL [20]
Figure 4: : User click on this URL which sets the variable pos to value of context i.e.Mary [20]
8
Example of the same URL with embedded malicious script.
Figure 5: An attacker embeds a malicious script as value of context field [20]
The user clicks on the above URL, which sends the request with the context value
as malicious JavaScript. The browser builds the DOM of the web page after receiving
the response from the server and sets the value of the property document.url to the
value of the context. When the script gets executed it updates the raw HTML of the
page with the malicious script and the malicious script now gets carried out by the
browser resulting in the attack.
Figure 6: HTML page with embedded malicious JavaScript [20]
2.2 Other variants of JavaScript Attack
Cross-site request forgery is also a standard JavaScript attack and has been listed
as number five in the Top 10 web applications security risks by OWSAP 2013 [2].
Cross-site request forgery refers to sending malicious requests to an authorized user
of websites that websites trusts. In cross-site request forgery, an attacker attempts to
send a state change request such as a fund transfer or an email change. An attacker
convinces an authorized user to execute unauthorized commands by use of social
engineering tricks such as sending an email that looks authorized to the user. By
9
clicking on the link may submit that forged request if the user is already login to the
website. A website has no way to know if the request is a legitimate one or a forged
one as a website stored the login credentials and other sensitive information of the
user in the cookies or session in the browser. That is why this attack is also known
as session over-riding attack.
A legitimate request example:
Alice wants to transfer funds to Bob’ account.
Figure 7: A legitimate fund transfer request to transfer money to Bob’s account usingGET request [20]
A malicious request
The attacker can change the value in GET request so that it transfers the fund
to the attacker’s account and tricks the victim using social engineering to click on
the below link to transfer money to his account. The below forged requests can be
by sent an email or can be injected in a website the user is most likely to visit while
transferring funds.
Figure 8: Malicious request forged by the attacker. Here name value is changed toMARIA form BOB and amount value is changed to 100000 form 100 [20]
10
2.3 Security measures adopted to prevent malicious JavaScript Attack
To avoid an attack, the following actions can be taken: escape and sanitize all the
users input data, whitelist input validation, and employ content security policy using
sandboxing. Modern web browsers are using the sandboxing and the same origin
policy to prevent or restrain a JavaScript attack: [5]. Sandboxing limits the scope of
a script, preventing the attacks from spreading system wide. The same origin policy
prevents a script from one source to access resources from a different origin. However,
attackers leverage the flaws in the websites and insecure practices and allowing them
to circumvent the above two restrictions. The common defects and unsafe practices
used by the attackers are vulnerable JavaScript inclusion and insecure JavaScript
generation [15]. JavaScript inclusion injects the third domain JavaScript in the top
level document by using the src attribute of a script tag and thus defy the purpose
of same origin policy [15]. Attackers use eval() function for dynamic generation of
malicious JavaScript code. According to research by [15], 66.4% of the website uses
the insecure practice of JavaScript inclusion, and 74.9% uses dynamic JavaScript
generation.
Modern approaches are using machine learning technology in combination with
de-obfuscation/emulation for better performance and accuracy [1]. Machine learning
can be used in analyzing and capturing the structural information of a malicious
JavaScript program by extracting the abstract syntax tree, while emulation can be
used to analyze the behavior of a malicious JavaScript program. Obtaining structural
information for analysis is known as static analysis while using emulation to execute
the exploit to examine and analyze the behavior and impact of an exploit is known
as dynamic analysis. According to TARDIS [1], dynamic analysis tends to be more
accurate than static analysis, but it has more performance overhead.
11
2.4 Static Analysis
Static analysis analyzes source code without executing it, and is commonly used
as a technique for troubleshooting a computer program [26]. Static analysis helps in
understanding the composition of a program. The static analysis examine weather a
software application is correct and consistent in its organization and depiction [26].
It can be performed automatically using specific tools such as parsers, data flow
analyzers, syntax analyzers, etc. Static analysis can also be followed by dynamic
analysis for uncovering the subtle defects or vulnerabilities. Static and dynamic
analysis together refers as glass box testing.
TARDIS is based on purely static analysis of malicious and benign JavaScript,
and combines static analysis with SLM for robust feature extractions. In our project,
we are using static analysis for analyzing the program syntax, and a JavaScript parser
for capturing the abstract syntax tree, and examining the structure and usage of
individual JavaScript statements, keywords, and reserved words. We are performing
automatic static analysis by parsing the scripts in the add-on. A more detailed
description of TARDIS is available in section 2.7.
2.5 Dynamic Analysis
Dynamic analysis involves examining source code by execution. It analyzes the
action, impact, and behavior of software before and after the execution of the software
in a controlled manner and environment. The execution of software can be carried out
in either artificial or real application environment. Path testing and branch testing
are two primary dynamic analysis techniques. Branch testing aims at traversing every
branch of a program at least once while path testing attempts to exercise as many
logical paths as possible [26].
12
Dynamic analysis and detection of a JavaScript exploit require a detection system
that can observe and examine the execution of a JavaScript code during run-time.
To capture this information, a JavaScript program is either executed in a sandbox
environment or the detection system interacts with the JavaScript engine of the web
browser. The detection system monitors and tracks the flow of the execution events,
which result in modifications to the environment state [26].
2.6 Related Work
This section presents the recently advanced approaches in detecting and analyz-
ing malicious JavaScript using machine learning technology. These approaches are
either using static analysis, dynamic analysis or a combination of both.
2.6.1 JStill (Mostly Static Approach)
The JStill [6] approach is static. However, in conjunction with static analysis,
JStill uses a lightweight runtime inspection, which helps in analyzing the essential
characteristics of an obfuscated malicious program. JStill performs static analysis to
capture the characteristics of an exploit. However, a static analysis alone may not
be accurate due to the obscured nature of the malicious program. An obfuscated
malicious program needs to be de-obfuscated before fulfilling its malicious intent and
requires particular function invocations. JStill leverages this observation of function
invocation to inspect the runtime behavior of obfuscated code. JStill examines the
function invocation pattern by a malicious program using the browser’s runtime oper-
ations and hence does not incur any extra performance overhead of dynamic analysis
that requires executing an exploit in a controlled environment. JStill can be imple-
mented in a browser. The average performance overhead of JStill is 4.9%. It shows
13
higher performance overhead i.e. > 8% for yahoo.com and sina.com.cn. JStill also
tends to give a higher false positive rate for a benign obfuscated JavaScript program.
2.6.2 Zozzle: Fast and Precise In-Browser JavaScript Malware Detection
Zozzle [24] is a combination of both static and dynamic analysis. Zozzle mostly
uses static analysis for better performance and high throughput. It also uses a com-
ponent of dynamic analysis for better accuracy and the analysis of an obfuscated
malicious JavaScript program. Static analysis of Zozzle uses Bayesian classification
and it uses the JavaScript abstract syntax tree’s hierarchical features to extract the
essential predictive features and quick scanning. To handle the obfuscation, Zozzle
uses a small runtime component. This component extracts and processes the JavaS-
cript that is generated at runtime using eval(), document.write(), etc. It then
sends this runtime generated code to its static analyzer right before the execution.
Zozzle has a very high throughput as big as one megabyte of JavaScript code per
second and an exceptionally low false positive rate of 0.0003%.
2.6.3 Cujo: efficient detection and prevention of drive-by-download at-tacks
Cujo [23] combined both static analysis and dynamic analysis for automatic de-
tection and blocking of drive-by download attacks. Static analysis extracts lexical
tokens representing reserved words, literals, and identifiers. The dynamic analysis
uses a lightweight sandboxing environment that analyzes execution behaviors. Both
the static and dynamic features are explained further using machine learning tech-
nique for robust detection of an exploit. Cujo can be embedded in a web proxy, and
it tends to reach a very high accuracy of 94% in detecting an attack with a very low
false positive rate. Cujo is a learning-based detection tool and uses the support vector
14
machine learning algorithm. In spite of high precision, the dynamic analysis part of
Cujo incurs performance overhead and the run time of Cujo is 500 ms per web page.
2.6.4 EarlyBird: Early Detection of Malicious Behavior in JavaScriptCode
EarlyBird [25] uses dynamic analysis to perform dynamic, efficient detection of
an exploit. A dynamic analysis requires execution of an exploit which may also result
in potential damage to the underlying system. EarlyBird attempts to prevent the
severity of harm caused by the execution of a malicious script by detecting it in on
early phase of execution. It uses a set of predefined events and JavaScript execution
results in particular sequences of these events. These event tracking can be used
for various features extractions. This sequence of events is then mapped to vector
space and uses linear support vector machine algorithm for learning and detection to
achieve better protection of the underlying system. EarlyBird restricts the amount
of exploit code that gets through the execution by a factor of 2. EarlyBird makes use
of support vector machine and can achieve a good performance of 93% with very low
false positive.
2.6.5 Prophiler: A Fast Filter for the Large-Scale Detection of MaliciousWeb Pages
Prophiler [16] uses static analysis for rapid detection of the presence of an exploit
in a web page. Prophiler uses a JavaScript program to extract significant features from
HTML content of a webpage. These features are then supplied to a machine learning
technology. The primary purpose of Prophiler is to reduce the resources and cost
of dynamic analysis tools for detection and analysis of a drive-by download attack.
Dynamic analysis tools are capable of detecting a drive-by download attack precisely,
15
but they have costly analysis. This overhead is generally too costly for performing
analysis on an extensive set of web pages. Prophiler is effective in reducing the load
of dynamic analysis tools by 85%, but it still incurred 270 ms per page and has a
13.7% false positive rate.
2.6.6 Wepawet
Wepawet [8] uses an emulation techniques and combines it with anomaly de-
tection for automatic identification of a drive-by download attack. Wepawet supplies
the features of regular JavaScript to the machine learning classifier and uses emula-
tion to detect the behavior of malicious anomalous JavaScript by analyzing it against
previously verified features. Wepawet achieves a low false negative rate and no false
positives on the data set tested.
2.6.7 PJScan: Static Detection of Malicious JavaScript-Bearing PDFDocuments [4]
A pdf document is a commonly used file format, and they provide many features.
Attackers have discovered a way to hide malicious scripts inside PDF files. PJScan
uses static analysis on extracted JavaScript code to detect the JavaScript-bearing
malicious PDF documents. PJScan incurs a significant low run-time overhead as
compared to other previous work done that uses dynamic analysis approaches. PJS-
can can work efficiently on both known and unknown malicious JavaScript. PJScan
utilized a lexical analysis approach and machine learning technology for automatic
construction of the models, which can then be used to detect a pdf attack.
16
2.6.8 IceShield: Detection and Mitigation of Malicious Websites with aFrozen DOM
IceShield [7] performs in browser dynamic analysis and de-obfuscation to detect
and mitigate a malicious JavaScript attack. IceShield is entirely based on dynamic
analysis. It de-obfuscates the code first and then performs analysis on an exploit
presented in clear text after de-obfuscation. IceShield primarily targets the types of
attack that compromise the DOM and injects malicious code. IceShield makes use of
a heuristic approach to discover an attacker from a benign user visiting and accessing
the web page. IceShield can identify the fragment of the webpage that is malicious
and modifies the page accordingly to block the attack. It is entirely implemented
in JavaScript, and hence lightweight. It is also independent of a browser and can
be applied in embedded browsers such as smartphone browsers. IceShield detection
accuracy is 98%, and performance overhead is 12ms for a website and 80 ms for a
smartphone.
Dynamic analysis provides better accuracy in detecting an exploit as compared
to static analysis, but it incurs a performance overhead. Static analysis is faster
than dynamic analysis, but not capable of detecting obfuscated malicious JavaScript
efficiently. After examining the recent works done towards the detection of mali-
cious JavaScript, we discover that most of the works are taking advantage of both
approaches. They are trying to be mostly static to achieve the desired speed and
implementing a lightweight dynamic analysis component for effectiveness without
sacrificing performance.
17
A snapshot of a benign JavaScript program
Figure 9: A sample of benign script form test data set
18
A snapshot of a malicious JavaScript program
Figure 10: A sample of malicious script form test data set
19
A snapshot of a obfuscated JavaScript program
Figure 11: A sample of obfuscated script form test data set
20
2.7 TARDIS
TARDIS (Towards Robust Detection of Malicious JavaScript) [1] developed by
Professor E J Jung et al. at the University of San Francisco, is a completely static
analysis tool. It only requires the source code of the exploit and hence does not require
execution and thus avoids dynamic analysis performance overhead. Text based static
analysis is not very useful in detecting obfuscated code as static analysis approaches
tend to have a high false positive rate on minified, obfuscated benign scripts. To
achieve optimal accuracy TARDIS has been supplemented with a powerful Statistical
Language Model.
TARDIS’s static analysis focuses on features that can differentiate between ma-
licious and benign scripts based on their textual attributes. Analyzing textual at-
tributes is purely static and does not require the execution of the source code. Some
example of these textual attributes can be the use of whitespace, line breaks, the
length of sentences, comments in a benign and malicious script, and the use of vari-
ous keywords. These textual attributes can be used to discover a pattern in the way a
malicious and benign JavaScript is written. These features alone are not sufficient for
detecting a malicious code efficiently. An attacker may avoid detections by a slight
change in their code generation algorithm, which requires analyzing more robust fea-
tures incurring significant work on the part of the attacker in modifying their code
generation algorithm to escape detection.
To achieve this requirement TARDIS makes use of a statistical language model
for automated feature extraction by using a JavaScript parser and an abstract syntax
tree in addition to the textual attributes features discussed in the previous section.
21
2.7.1 Abstract syntax tree
An abstract syntax tree defines the syntactical structure of a program by using
nodes of a tree. An AST represents constants or variables as leaf nodes, and operators
and statements as an inner node of the AST. Characteristic of an abstract syntax
tree can be used to extract features that are difficult to be evaded by an attacker.
Modification in the features of AST towards avoiding detection will require imitation
of the AST of a benign code. A malicious code makes use of certain functions with
higher frequency to carry out attacks such as string concatenation or fromCharCode
() etc. Concealing the detection of these features by an AST will require the attacker
to use a new algorithm to generate malicious code that avoids the textual attributes
detection [1].
2.7.2 Statistical Language Modeling (SLM)
Statistical language modeling (SLM) [11] makes use of a statistical language
model. A statistical Language model is defined as a probability distribution of a
string (s) in a sentence [11]. The probability distribution of a string (s) represents the
frequency of occurrence of (s) as a sentence. The most widely used SLM techniques
are N-gram models and its variants [11].
TARDIS makes use of SLM for automatic feature extraction by employing a
JavaScript parser. The JavaScript parser parses benign and malicious scripts and
extracts essential features. These extracted features are then used to create SLM
benign and SLM malicious training model that can be used to classify a benign or a
malicious script.
22
2.7.3 TARDIS SLM model
The parser generates a collection of words based on certain delimiters after pars-
ing a training corpus. These words then can be appended together and form an
n-gram. N-gram represents a consecutive sequence of n words from a sentence. These
n-grams constitute the features of the training model. The SLM training model de-
scribes the features as key-value pairs, where the key denotes a feature/n-gram and
the values represents the probability of occurrence of that particular element in the
model. This mapping of n-grams with probability forms the statistical model of
TARDIS’s static analysis technique. This mapping can then be used to compute the
probability that a document belongs to a particular class (benign or malicious) [1].
TARDIS generates SLM models for benign and malicious scripts. SLM benign
models are computed over benign scripts while the SLM malicious models are calcu-
lated using malicious scripts. While testing both the models are used to estimate the
overall probability of a document belonging to either of the models. The model that
gives the higher probability wins and the testing script is classified to the winning
model.
TARDIS makes use of the following formula to estimate the likelihood of cat-
egorization of a script to either the benign or the malicious category.
2.7.4 N-grams SLM model
An n-gram model can have different forms, and each of these forms can be used
in generating a model. Each of these models can provide different information and
as well as the features and can have a different impact on the words and probability
mapping, precision of the model. TARDIS experimented with models computed based
on n-grams of size one, two, three, and four to tune the accuracy. N-grams of size
23
one considers each character as a feature while n-grams of size two joins together two
consecutive characters. Similarly, a model based on n-grams of size three and four
can be computed. N-grams model of size one tends to lose the surrounding context
while n-grams model of a large size can provide too many surrounding contexts but
less meaningful matches [1]. Mostly n-grams of size two or three provide meaning
full match with adequate surrounding contexts. TARDIS built its training model for
n-grams of size one to n-grams of size four and compute the accuracy of each of the
model in order to identify which n-grams model provides better accuracy in terms of
classification. TARDIS proposes the use of three categories of n-gram model. Each
of them computes the benign and malicious training model for n-gram of size one,
two, three, and four.
2.7.5 Character level n-grams
According to TARDIS, a character level n-grams model expresses the content of
an input script rather than the composition of the input script. A character level n-
grams model uses characters as tokens. It converts the input sequence to a collection
of the characters and joins the consecutive characters to form different sizes of n-
grams.
Given a sequence of input script as
var str = "javaScript"
An n-gram of size one will look like
[’v’, ’a’, ’r’, ’ ’, ’s’, ’t’, ’r’, ’=’, ’"’, ’j’, ’a’, ’v’, ’a’,’S’, ’c’, ’r’, ’i’, ’p’, ’t’, ’"’]
An n-gram of size three will look like
24
[’v’, ’a’, ’r’], [’a’, ’r’, ’ ’], [ ’r’, ’ ’, ’s’], [’ ’, ’s’, ’t’],[’s’, ’t’, ’r’], [’t’, ’r’, ’=’], [’r’, ’=’, ’"’], [’=’, ’"’, ’J’],[’"’, ’J’, ’a’], [’a’, ’v’, ’a’], [’’v’, ’a’, ’s’], [’a’, ’S’, ’c’],[’a’, ’S’, ’c’], [’S’, ’c’, ’r’], [’c’, ’r’, ’i’], [’r’, ’i’, ’p’],[’i’, ’p’, ’t’], [’p’, ’t’, ’i’]
A character level n-grams model can successfully extract useful predictive fea-
tures such as JavaScript keywords, operators, and frequency of use of increment,
decrement operators, etc. However, it is not very informative regarding the structure,
and semantically meaningful input sequences such as function call as a character level
n-grams model break down the function call into a list of characters.
2.7.6 Keyword Transformation
Keyword transformation n-grams model reserves all the JavaScript keywords as
they are and uses them without breaking down into character tokens. It treats all the
other input sequence the same as character level n-grams and calculates the model for
different n-gram size. TARDIS uses a list of reserved JavaScript keywords to identify
the keywords in an input script. Keyword transformation also does not count space
character in the model generation.
Given a sequence of input script as
var s = 10;
Keyword transformation n-grams of size one will look like
[’var’, ’s’, ’=’, ’1’, ’0’, ’;’]
Keyword transformation n-grams of size three will look like
[’var’, ’s’, ’=’], [’i’, ’=’, ’1’], [’=’, ’1’, ’0’], [’1’, ’0’, ’;’]
25
Here ’var’ is a JavaScript keyword and hence, it is used as it is without breaking
down into characters. Keyword transformation represents both the semantics and the
content of a program. Keyword transformation can be used in extracting common
programming language features such as variable assignments, which is helpful in clas-
sifying a benign script if it is not obfuscated [1]. However, it does not prove very
beneficial in identifying malicious, obfuscated scripts [2].
2.7.7 Composite word-type transformation
Keyword transformation is not very accurate in analyzing obfuscated JavaScript.
An obfuscated JavaScript program makes use of string encoding to conceal its payload.
Keyword or character level conversion on an encoded string results in a substantial
number of unique characters that do not present any significant information. To
manage efficient detection of obfuscated malicious JavaScript, TARDIS is uses com-
posite word type transformation. The composite word type transformation practices
a predefined class based transformation. It assigns each token to a particular class
and computes the probability model by computing the frequency of appearance of
these classes in the model. Representing a program based on these classes reduces
randomness in a program to more significant features. Commonly a program consists
of digits, hexadecimal numbers, white spaces, punctuation, etc. Composite word type
transformation provides a separate class for each type of element. Characters other
than the above-defined classes are combined and interpreted as whole words.
Composite word type transformation n-grams of size one of ’var s = 10;’
[’var’, ’SPACE’, ’s’, ’SPACE’, ’PUNCTUATION’, ’SPACE’,’DIGIT’, ’PUNCTUATION’]
26
Composite word type transformation n-grams of size three of ’var i = 3 ;’
[’var’, ’SPACE’ , ’s’], [’SPACE’ , ’s’, ’SPACE’],[ ’s’, ’SPACE’,’PUNCTUATION’], [’SPACE’,’PUNCTUATION’,’SPACE’],[’PUNCTUATION’,’SPACE’, ’DIGIT’], [’SPACE’, ’DIGIT’, ’PUNCTUATION’]
2.8 Malicious Probability Query Strategy
A composite word type transformation reduces randomness and uniqueness of
an obfuscated JavaScript program and group together the unique characters using
a predefined class. Probability model generation of an obfuscated script requires
extra control over the method by which probability of a particular type of n-gram is
estimated. TARDIS introduces an alphanumeric probability strategy for computation
of malicious model. An alphanumeric probability strategy calculates the probability
of string consists of only alphanumeric characters based on the following formula
(1/62)𝑛
where n is the length of the string. Here 62 is the sum of 26 upper case alphabets
from A to Z, 26 lower case alphabets from a to z, and ten digits from 0 to 9.
TARDIS also performs smothering of the probability of an n-gram which is not
present in the model to avoid setting the probability as zero.
27
CHAPTER 3
Firefox add-on Implementation
Firefox add-ons are a small piece of software that are used to extend and modify
the installed version of Firefox by adding new features or functionality. An add-on can
be used to change the theme or visual appearance of a website, add new features to the
installed Firefox version, modify the user interface, add foreign language dictionaries,
etc. Standard web technologies such as JavaScript, HTML and CSS are commonly
used to develop a Firefox add-on [18].
3.1 Usability of our Firefox add-on
We have developed a Firefox add-on to integrate TARDIS with the web browser.
It scans the JavaScript from the currently open tab and alerts the user to the pres-
ence of a malicious script, hence preventing the user from any further action in the
currently open tab. The central purpose of developing a Firefox add-on is to show
the usability and performance evaluation of TARDIS in the browser. The add-on is
entirely developed in JavaScript and hence can be integrated with other analysis tools
in JavaScript.
3.2 Developing a Firefox add-on
A Firefox add-on can be developed using either of the following two methods:
1. WebExtensions
2. Add-on SDK
28
3.3 WebExtensions
WebExtensions provide APIs for developing Firefox add-on, and is currently
in the early state, but is considered to be the future of Firefox add-on development.
According to [18], WebExtensions will become the standard by 2017. WebExtensions
provide cross-browser compatibility, and the APIs are compatible with Google chrome
and Opera’s Extension API [18].
3.4 Add-on SDK
The add-on SDK method provides JavaScript APIs for Firefox add-on devel-
opment and tools for creating, running, testing, and packaging them. Standard web
technologies (JavaScript, CSS, HTML) are used in combination with the add-on SDK
APIs. It requires Firefox version 38 or later [18].
We have developed our Firefox add-on using the add-on SDK. At the time we
started development, add-on SDK was the most stable version available.
3.5 Firefox Add-on SDK installation and structure
The add-on SDK includes the jpm for initializing, running, testing, and packaging
a Firefox add-on. jpm is based on Node.js. After installation, an empty add-on is
initialized by running ’jpm init’ inside an empty directory. The initial directory
structure of a Firefox add-on looks like the following:
The figure shows the directory structure of the add-on. Here index.js is the
entry point of the add-on and can be changed during the initial setup. Once the
initial setup is done, Firefox add-on is developed using Add-on SDK’s high-level and
low-level APIs.
29
Figure 12: Initial directory structure of the Firefox add-on [18]
3.6 index.js
Figure 13: Index.js
30
Index.js is the entry point of our Firefox add-on. Index.js creates and adds a
button to the current version of Firefox. On the onClick event of the add-on button,
function runScript gets invoked. The runScript function is responsible for invoking
the SLM_Script.js file and including the pre-build training models.
Index.js is our main add-on script. An add-on scripts can use the SDK’s high-
level and low-level APIs. But it does not get access to the web content directly. The
add-on uses separate scripts known as content scripts to get access to the web content.
To scan the JavaScript present on the page and detect malicious content, our add-on
needs to access the web page content. Some of the SDK API’s, like page-mod and
tabs, provide necessary functions to load content-script. Here we are loading content
scripts in our main SDK script using the tabs module’s attach function. The attach
function is using the contentScriptFile option to load content script as a file.
Figure 14: function runscript
Tabs module is using attach () function to load the content scripts.
Self.data.url(file_name) is pointing to the file inside data directory.
3.7 Content scripts
Content scripts can access web content, but like the main add-on scripts, content-
scripts can’t access the SDK’s APIs. Content scripts are stored as separate files under
31
the data directory. The data directory is not created by default and needed to be
added manually. We store all of our content scripts and a precompiled training model
inside the data directory. The content script can communicate back its response to
the add-on script using message passing APIs.
The message communication can be done using the property port of the global
object self. The sender the of message calls port.emit to send message and the
receiver calls port.on to receive the message.
3.8 Data Directory
The data directory contains the necessary content scripts that extracts the scripts
from the web page of the current open tab and classify them as either benign or
malicious category.
32
Figure 15: Add-on directory structure. Data directory contains models, image, andcontent scripts.
33
3.8.1 SLM_Script.js
SLM_Script.js is a content script. SLM_Script.js extracts the JavaScript from
the web page and stores it in an array and then applies algorithm to automatically
generate the n-gram based benign and malicious SLM models. The script can gener-
ate the following SLM models: character level n-grams of size three and four, keyword
transformation n-grams of size three and four, and composite word type transform-
ation n-grams of size three and four. These features are used by the precompiled
benign and malicious training models to compute the overall probability of the script
belonging to either of the models. The result is then passed to the add-on script
index.js using port.emit.
1 self.port.emit("script -response", "compositeMaliciousScore = "+ compositeMaliciousScore);
2 self.port.emit("script -response", "compositeBenignScore = " +compositeBenignScore);
Figure 16: Example of port.emit: SLM_scripts.js passing the final result to theindex.js
3.8.2 Models
The Firefox add-on leverages the benefit of a pre-compiled training model for
detection efficiency and better performance. The models directory inside the data
directory holds all the precompiled training models required by the add-on. A script
is tested on both the training model to detect the presence of malicious content. A
precompiled model used within the Firefox add-on saves the overhead of sending and
receiving a HTTP request to the server for the classification decision.
34
3.9 Pre-compiled training model
This section will present the detail discussion of the pre-compiled models we are
utilizing for the add-on.
3.9.1 Types of pre-compiled models
We categorize all the training models to two categories: benign and malicious
Each of the benign and malicious categories further contains models based on charac-
ter level n-grams, keyword transformation n-grams, and composite word type trans-
formation n-grams. We are computing n-grams models of each type of size three and
four.
35
3.9.2 Character level n-gram model
To compute a character level n-gram model, a file is parsed and then converted
to a list of characters, then consecutive characters are joined and stored as a key-value
pair in JSON format. A key is the n-gram/feature and the value is the frequency of
occurrence in the script. This type of model presents the content of the document
more than the structure.
Figure 17: A snapshot of a pre-compiled malicious character level n-grams model ofsize four. Every key is four characters long
36
3.9.3 Keyword transformation
Keyword transformation parse the script and converted it into a list of characters,
then join the consecutive characters to form n-grams. Keyword transformation is
similar to character level n-grams, but in keyword transformation, reserved keywords
are stored with the whole word as a single token. Keyword transformation preserves
both the content and the semantics of a script.
Figure 18: a snapshot of a keyword transformation n-grams model. Reserved keywordsuch as length, constructor, and min appear as the whole word combined with theconsecutive characters that are not part of the reserved keyword
37
3.9.4 Composite word type transformation
Composite word type transformation converts the sequence of characters into
distinct classes. Here the following classes are used to represent characters: DI-
GIT, HEX, WHITESPACE, PUNCTUATION, and PERCENT. Characters other
than these categories are joined and represent a single token. These classes and
tokens are combined to form composite word type n-grams of size there and four. As
discussed in the section 2.7.4, composite word type transformation reduces entropy
in an obfuscated malicious program.
Figure 19: A snapshot of the malicious n-grams composite word type transformationmodel
38
3.9.5 Precompiled training models computation
The models are computed using the TARDIS source program in Java. TARDIS
is written in Java. The source code first computes the training model and uses the
training model to test the JavaScript for malicious or benign categorization. We
leverage this functionality and store the model generated by TARDIS persistently in
JSON format. The primary reason behind storing a model in JSON format is that a
JSON object is lightweight and portable. Storing a model in JSON with the add-on
would not take much space in the browser and it can also provide a quick look up of
key-value pair.
39
1 JSONObject obj_benignCountsAB = new JSONObject ();2 for(Map.Entry <TermSequence , Integer > entry :
benignModel.countsAB.entrySet ()){3 String key = entry.getKey ().toString ();4 key = key.substring (1, key.length () -1).replace(", ",
"");5 int value = entry.getValue ();6
7 try {8 //if(value > 10)9 obj_benignCountsAB.put(key , value);
10 } catch (JSONException e) {11 e.printStackTrace ();12 }13 }14
15 FileWriter file_benignCountsAB;16 try {17 file_benignCountsAB = new
FileWriter("KeywordBenignCountsAB_50.json");18
file_benignCountsAB.write(obj_benignCountsAB.toString ());19 file_benignCountsAB.flush();20 file_benignCountsAB.close();21 } catch (IOException e2) {22 e2.printStackTrace ();23 }
Figure 20: Java code added for model computation
3.9.6 Problems faced during pre-compiled model generation and solutionimplementation
The model generation for large no of files is a computationally expensive process.
For efficient processing and time reduction for model generation, we implemented a
multithreading solution to the existing TARDIS model generation algorithm. The
multithreading solution reduces execution time by roughly two-third.
40
Figure 21: output of top command before multithreading implementation shows %CPU utilization as 99.7%
Figure 22: CPU idle time before multithreading implementation = 92
CPU utilization percentage and idle time after multithreading implementation.
41
Figure 23: output of top command after multithreading implementation shows %CPU utilization as 346.2%
Figure 24: CPU idle time after multithreading implementation = 76
42
3.10 Firefox add-on implementation
After the installation, the Firefox add-on appears in the browser toolbar on the
right side. A user can click on the add-on to perform malicious scripts detection. The
add-on extracts the JavaScript from the page and parses the script. After parsing,
the add-on calculates the probability score of the script and classify it either benign
and malicious based on the score. The add-on then passes the result to console using
port.emit.
Figure 25: Firefox add-on in the browser
Figure 26: Firefox add-on detection result in the console
43
3.11 Result Computation
The add-on computes the overall probability of a script over the benign and
malicious model. For each n-gram the mode looks for the frequency value in n-grams
of size three model and n-grams of size four JSON model. The model then computes
the overall probability of the script for both the benign and malicious models using
the formula [1]
𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 = 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 +𝑚𝑎𝑡ℎ.𝑙𝑜𝑔(𝑝𝐴𝐵/𝑝𝐴)
𝑝𝐴𝐵 = (𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 𝑛− 𝑔𝑟𝑎𝑚𝑠 𝑜𝑓 𝑠𝑖𝑧𝑒 𝑓𝑜𝑢𝑟)/(𝑡𝑜𝑡𝑎𝑙 𝑛𝑜 𝑜𝑓 𝑤𝑜𝑟𝑑𝑠)
𝑝𝐴 = (𝑓𝑟𝑒𝑞𝑢𝑒𝑛𝑐𝑦 𝑜𝑓 𝑡ℎ𝑒 𝑛− 𝑔𝑟𝑎𝑚𝑠 𝑜𝑓 𝑠𝑖𝑧𝑒 𝑡ℎ𝑟𝑒𝑒)/(𝑡𝑜𝑡𝑎𝑙 𝑛𝑜 𝑜𝑓 𝑤𝑜𝑟𝑑𝑠)
44
CHAPTER 4
Testing
4.1 Dataset
We have obtained dataset for our model computation from different sources. We
have collected a significant amount of both malicious and benign scripts to train our
model, and we have made the effort to include various types of malicious scripts
such as redirection, obfuscation, etc. For benign scripts set, we have also considered
minified obfuscated benign scripts.
4.1.1 Malicious scripts
we have collected over 50000 of malicious scripts from EJ Jung et al. and the
research team from the University of San Francisco. To train our model, we are
utilizing 15000 of malicious datasets of a size of total 200 megabytes. Half of the
malicious scripts is of type redirection, and other half represents all the other forms
of attack.
4.1.2 Benign Scripts
We have collected the benign scripts from various resources on the internet. We
have obtained over 27000 of benign files of total size equal to 200 megabytes. These
files represent both clear and obfuscated benign scripts. Most of the benign files are
from the JavaScript libraries such as React.js, MooTools, JQuery, D3.js, Processing.js,
etc.
45
4.1.3 Problems with the scripts
In out dataset, it has been observed that malicious script size is commonly bigger
than the benign script size. To match the different size, we are using maximum 15000
files for malicious model computation and over 30000 for benign model computation.
We have also made sure that that both the models are of equal size to avoid over-
fitting.
4.2 Training models
We are testing the add-on for various size of the models. We have observed that
while calculating models if we optimize the model and don’t consider the n-grams
with the frequency less than 10, the model size gets reduced significantly. However,
this reduction in size may incur a loss in accuracy. We have tested the add-on for
both optimized and non-optimized version of each type of transformation. We are
capturing accuracy and detection time with the different size of the models of each
category to identify the maximum size of the training model that the add-on can
utilize without sacrificing the performance.
We have computed benign and malicious models for a total file size of 50 mega-
bytes for all the three kinds of transformation: character level n-gram, keyword trans-
formation, and composite word type transformation. A detailed description of these
transformation can be found in section 2.7.4.
4.3 Evaluation of n-grams models
We have evaluated all the three models for accuracy and performance. This
section describes in details the performance and accuracy trade-off in between the
models.
46
4.3.1 Evaluation of optimized and non-optimized models of characterlevel n-grams.
We have evaluated all the three models for accuracy and performance. This
section describes in details the performance and accuracy trade-off in between the
models.
Table 1: Performance comparison of optimized and non-optimized character leveln-grams model.
Model size(mb) Scripts TP FP TN FN Accuracy Detectiontime/-file
Character(opt) 8 3574 1299 164 1630 481 82% 150msCharacter(non-opt)
33.6 3618 1430 170 1624 394 85% 507ms
We observe that the no of scripts computed are changing with the change in
models. If the model calculates the probability of a script as 0, it does not take that
particular script under consideration.
Table 2: Accuracy and precision evaluation of optimized and non-optimized characterlevel n-grams model.
Model Accuracy Malicious Precision Benign PrecisionCharacter(opt) 82% 41% 59%Character(non-opt) 85% 44% 56%
A non-optimized character level n-grams model has 85% accuracy on the sample
dataset as compared to the optimized character level n-grams model. However, there
is a trade off in between benign and malicious precision and size of the model. A non-
optimized model can identify a malicious script with high precision but the model
size is 4 times bigger as compared to optimized one.
47
4.3.2 Keyword transformation
Table 3: Performance comparison of optimized and non-optimized Keyword trans-formation n-grams model.
Model size(mb) Scripts TP FP TN FN Accuracy Detectiontime
Keyword(opt) 7.31 3513 638 0 1767 1108 68.4% 5105msKeyword(non-opt)
31.6 3476 665 0 1767 1044 70% 2217ms
Table 4: Accuracy and precision evaluation of optimized and non-optimized keywordtransformation level n-grams model.
Model Accuracy Malicious Precision Benign PrecisionKeyword(opt) 68.4% 18% 82%Keyword(non-opt) 70% 19.1% 80.9%
The keyword transformation models achieve notable low accuracy as compared
to the other two models. The accuracy improves for the non-optimized version of the
model. We also observe that the models’s malicious precision is very high as compared
to model’s benign precision. The low accuracy and precision of the model can be
attributed to the large difference in the total no of words in keyword transformation
benign models and the keyword transformation malicious model. The reason of such
big difference can be that a malicious obfuscated models may have more random
strings and less reserved keywords as compared to the benign model of same size.
48
Figure 27: Total number of words in benign keyword transform model
Figure 28: Total number of words in malicious keyword transform model
4.3.3 Composite word type transformation
Table 5: Evaluation of composite word type transformation n-grams model.
Model size(mb) Scripts TP FP TN FN Accuracy Detectiontime
Composite(opt) 11 3544 1687 9 1758 90 97.2% 1000ms
Table 6: Accuracy and precision evaluation of composite word type transformationn-grams model.
Model Accuracy Malicious Precision Benign PrecisionComposite(opt) 97.2% 47.9% 52.1%
The composite word type transformation model provides a very good accuracy
of 98.7%. compared to other two types n-grams model with reasonable performance.
The good performance of composite word type n-grams model is due to the reason that
it represents the characters in the script to the set of classes. A malicious obfuscated
script may contain random strings which may not represent the semantics of a script.
However, if these random strings are converted to some set of particular words, then
it provides more meaning to the script.
49
Here we have not considered the non-optimized model of the composite word
type transformation due to its large size i.e. 174.9 megabytes. A model of such
a large size is not optimal for a Firefox add-on in terms of performance and space
complexity. The composite word type transformation also achieves good malicious
and benign precision.
4.4 Model comparisons regarding accuracy and detection time on sampledata set.
Figure 29: Accuracy comparison of all the three models
The composite word type transformation n-grams model achieves the highest
accuracy compared to character level n-grams and keyword transformation n-grams
models on the sample dataset.
50
Figure 30: Performance comparison of all the three models
A character level n-grams model provides the lowest detection time in comparison
to the other two models. Keyword transformation performs worst.
51
Figure 31: Performance comparisons of all the three models in real word scenario forthe top websites
The composite word type transformation n-grams model provides an average
detection time of 2.7 second with very good accuracy of 98%. The character level
n-grams model performs best in terms of detection time. However, the character level
n-grams model does not achieve the best accuracy. Keyword transformation performs
worst in terms of accuracy and detection time. We also observe that both the keyword
and composite transformation give the worst performance for amazon.com.
52
CHAPTER 5
Summary
Our experiments shows that our Firefox add-on achieves a maximum accuracy
of 97.2% with the average detection time 1 s. We also observe that the compos-
ite word type transformation has better accuracy than character level and keyword
transformation n-grams models. However, there is a trade-off between the accuracy
and the performance in between the character level n-grams model and composite
transformation model. The composite word type transformation model achieves very
high accuracy but, it also requires high detection time for certain websites. Similarly,
the character level transformation provides accuracy of 85% but it achieves high per-
formance. The keyword transformation model performs the worst compared to the
other two models with regard to both the accuracy and detection time. To the best
of our knowledge, our add-on is the only one of its kind that is using a precompiled
training model stored in JSON format within add-on. The add-on achieves similar
accuracy to TARDIS. We have also observed that while computing the model and
converting to JSON, we are losing certain encoded data which may have a significant
effect on the accuracy of the model.
Given the size of the dataset and training model, our add-on achieves an ex-
cellent performance. However, attackers continuously find new and evolved method
to perform attacks. To incorporate the new features required for detection of the
evolving attacks, we need to keep on updating our precompiled model on a timely
basis. To do this, one could have a server which continuously collects new data set
and computes the new model and then updates the old one. In the current set-up, the
model computation is a computationally expensive- process. To improve this process,
53
we can leverage map-reduce or other similar technology that can handle large files in
small time.
54
LIST OF REFERENCES
[1] Peter Likarish, Eunjin (EJ) Jung, Chris Boyce. Towards a Robust Detection ofMalicious JavaScript(TARDIS), Unpublished paper
[2] The Open Web Application Security Project(OWASP). The owasp top ten pro-ject. https://www.owasp.org/index.php/Category: OWASP_Top_Ten_Project./
[3] Peter Likarish, Eunjin (EJ) Jung, Insoon Jo. Obfuscated Malicious JavaScriptDetection using Classification Techniques, second edition, Prentice Hall, 1991.
[4] Pavel Laskov and Nedim Srndic. Static Detection of Malicious-JavaScript-bearing PDF Documents. In Proceedings of the 27th Annual Computer SecurityApplications Conference. ACM, 2011.
[5] Same-origin policy. Web. Retrieved December 2015, fromhttps://developer.mozilla.org/en- US/docs/Web/Security/Same-origin_policy
[6] Wei Xu, Fangfang Zhang, Scencum Zhu. JStill: Mostly Static Detection of Ob-fuscated Malicious JavaScript Code. In Proceedings of the third ACM conferenceon Data and application security and privacy. ACM, 2013.
[7] Mario Heiderich, Tilman Frosch and Thorsten Holz. IceShield: Detection andMitigation of Malicious Websites with a Frozen DOM. Recent Advances in In-trusion Detection Volume 6961 of the series Lecture Notes in Computer Science.pp 281-300.
[8] Phu H. Phung, Lieven Desmet. A Two-tier sandbox architecture for untrustedJavaScript. In Proceedings of the Workshop on JavaScript Tools. Pages 1-10.ACM, 2012.
[9] Fergal Glynn. JavaScript Security. Web. Retrieved March 2015, fromhttp://www.veracode.com/security/javascript-security
[10] Kevin Whinnnery. Comparing Titanium and PhoneGap. (May 12, 2012). Web.Retrieved March 2015, fromhttp://www.appcelerator.com/blog/2012/05/comparing-titanium-and-phonegap/
[11] What is Statistical Language Modeling. Web. Retrieved April 2015, fromhttp://homepages.inf.ed.ac.uk/lzhang10/slm.html
[12] Cross Site Scripting. Retrieved April 2015, Web. Retrieved March 2015, fromhttps://www.owasp.org/index.php/Cross-site_Scripting_(XSS)
55
[13] Non- Persistent cross site scripting. Web. Retrieved April 2015, fromhttp://www.acunetix.com/blog/articles/non-persistent-xss/
[14] Client-Side Attacks. Web. Retrieved April 2015, fromhttp://neokobo.blogspot.com/2014/01/3218-client-side-attacks.html
[15] Chuan Yue, Haining Wang. Characterizing Insecure JavaScript Practices on theWeb. In proceedings of the 18th international conference on World wide web.pages 961-970. ACM 2009.
[16] Davide Canali, Marco Cova, Giovanni Vigna, Christopher Kruegel. Prophiler: AFast Filter for the Large-Scale Detection of Malicious Web Pages. In proceedingsof the 20th international conference on World wide web. pages 197-206. ACM2011
[17] Obfuscation of client side JavaScript. Web. Retrieved April 2015, fromhttp://stunnix.com/prod/jo/sample.shtml
[18] Firefox Add-ons SDK. Web. Retrieved April 2015, fromhttps://developer.mozilla.org/en-US/Add-ons
[19] Cross-Site Request Forgery (CSRF). Web. Retrieved April 2015, fromhttps://www.owasp.org/index.php/Cross-Site_Request_Forgery_(CSRF)
[20] DOM-based Cross-Site Scripting(XSS) explained. Web. Retrieved April 2015,fromhttp://www.acunetix.com/blog/articles/dom-xss-explained/
[21] ChengXiang Zhai. Statistical Language Models for Information Retrieval A Crit-ical Review. Web. Retrieved March 2016. Fromhttp://sifaka.cs.uiuc.edu/czhai/pub/slmir-now.pdf
[22] M. Stamp, Information Security: Principles and Practice, second edition, Wiley,2011
[23] Konrad Rieck, Tammo Krueger, Andreas Dewald. Cujo: Efficient Detection andPrevention of Drive-by-Download Attacks. In Proceedings of the 26th AnnualComputer Security Applications Conference. Pages 31-39. ACM, 2010
[24] Charlie Curtsinger, Benjamin Livshits, Benjamin Zorn, Christian Seifert.ZOZZLE: Fast and Precise In-Browser JavaScript Malware Detection. In Pro-ceedings of the 20th USENIX conference on Security. Pages 3-3. USENIX Asso-ciation 2011
[25] Kristof Schutt, Marius Kloft, Alexander Bikadorov, Konrad Rieck. Early Detec-tion of Malicious Behavior in JavaScript Code. In Proceedings of the 5th ACMworkshop on Security and artificial intelligence. Pages 15-24. ACM, 2012
56
[26] static analysis (static code analysis). Web. Retrieved April 2015, fromhttp://searchwindevelopment.techtarget.com/definition/static-analysis
[27] Glass box testing. Web. Retrieved April 2015, fromhttp://www.issco.unige.ch/en/research/projects/ewg95/node81.html
57
![Analyzing of Malicious JavaScript[en]](https://img.pdfslide.us/doc/110x75/58a28aa01a28ab891a8b7fe3/analyzing-of-malicious-javascripten.jpg)
